Study On The Face Recognition By Sparse Representation Algorithm In Reproducing Kernel Hilbert Space

Face recognition is a classical yet challenging research topic in computer vision and pattern recognition.At present,the Sparse Representation(SR)algorithm has been playing an important role in face recognition and has attracted lots of attention.Under the SR framework,the Sparse Representation based Classifier(SRC)and the Collaborative Representation based Classifier have been proposed and achieved superior performance to other methods,such as support vector machine(SVM)for face recognition.However,SRC or CRC considers the training samples in each class contributing equally to the dictionary in that class,i.e.,the dictionary consists of the training samples in that class.This may lead to high residual error and instability.Hence,Class specific dictionary learning(CSDL)for sparse representation and collaborative representation based classifier is proposed.Both CSDL-SRC and CSDL-CRC methods greatly reduce high residual error and instability and improve the recognition rate in face recognition.A basic assumption of above methods is that they are processed in the Euclidean space.In many practical applications,however,data often demonstrate intrinsic non-linear structures and relationships.In the thesis,in order to promote the ability of dealing with non-linear structures and relationships,the above algorithms are extended to the Reproducing Kernel Hilbert Space(RKHS).The proposed methods can efficiently capture the hidden non-linear information of complex data,reduce the feature reconstruction error and learn discriminative sparse codes for face recognition.The main contributions are threefold.First of all,we deeply analyze the theory and application of Sparse Representation(SR)algorithm and sum up their existing problems and shortcomings.On the other hand,the great deficiency that the SR algorithm cannot capture the intrinsic non-linear structures and relationships is pointed out.Secondly,Kernel Sparse Representation based Classifier(KSRC)and Kernel Collaborative Representation based Classifier(KCRC)are proposed.We give a brief introduction of SRC and CRC method,and then extend SRC and CRC into Reproductive Hilbert Kernel Space.Extensive experimental results show that the performance of KSRC and KCRC achieve superior performance for face recognition on several benchmark databases.Thirdly,Class specific dictionary learning(CSDL)for KSRC(CSDL-KSRC)and Class specific dictionary learning(CSDL)for KCRC(CSDL-KCRC)are proposed.First,an explicit description for Class Specific Dictionary Learning(CSDL)method is given.Then,we extend the CSDL to the Reproducing Hilbert Kernel Space and form the Class Specific Dictionary Learning for Kernel Sparse Representation based Classifier(CSDL-KSRC)and the Class Specific Dictionary Learning for Kernel Collaborative Representation based Classifier(CSDL-KCRC).Extensive experimental results show that the performance of CSDL-KSRC and CSDL-KCRC achieve superior performance for face recognition on several benchmark databases.